On the prediction of protein-ligand structural complexes and binding affinities by hybrid statistical scoring function

• Main Author: Oon, Yu Yang
• Other Authors: Mu Yuguang
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Science::Physics; Science::Biological sciences::Biophysics
• Online Access: https://hdl.handle.net/10356/166416
• Institution: Nanyang Technological University

Computer-aided drug discovery has truly revolutionised the way we think about and how we develop new drugs targeted to treat obnoxious diseases. Among all the computational methods, scoring functions play a fundamental role in virtual screening, in which we screen through large chemical databases to identify potential drug candidates. Due to the expeditious expansion of computational power, there has been a recent eruption in the invention and availability of scoring functions. However, these algorithms either lack of complexity to capture deeper insights into the chemical interactions between receptors and ligands to predict accurate binding affinities, or are simply “black boxes” like machine learning-based methods, with zero transparency to provide any interpretability on how predictions are made. Therefore, it is necessary to integrate the advantages of different types of classical scoring functions to create a hybrid scoring function that can exploit information about diverse ligands that have been recognised to bind to the same receptor. The strategy of exploiting the pattern of the formation of protein-ligand interactions from experimental measurements and the resemblance to the common scaffolds shared by these ligands, will point the way when medicinal chemists or pharmacologists are confronted with unknown hits compounds or drug targets. Our novel method — Re-ComBind, which is based on a statistical framework, leverages the advantages of its predecessor ComBind and employs Quick Vina 2 as the baseline per-ligand scoring function. Despite its lower performance in the screening power test, Re-ComBind has been proved to substantially improve the performance of its baseline function on a series of benchmarks. This study raises broad possibilities of improving the accuracy of predicting binding affinity by incorporating orthologous sources of information on ligands and acting as an additional correction term for bespoke machine learning-based scoring functions in the future.